Training workloads are highly variable and mostly consistent across the cluster (all systems ramp up and down at the same time).